System Config Keys Reference

The Platform exposes 71 runtime configuration keys through the admin API. These keys control gateway behavior — DLP thresholds, rate limits, audit settings, agent execution, quality scoring, and more — without requiring a service restart or redeployment.

Resolution order

Every configuration key resolves through a three-tier priority chain:

Database override — value set via the admin API (PUT /api/v1/admin/config/{key})
Environment variable — value from the process environment
Compiled default — hardcoded default in the application

Database overrides take highest precedence. When an admin updates a key via the API, the new value is persisted in the system_configs table and takes effect immediately for subsequent requests.

Admin API endpoints

All system config endpoints require admin-role JWT authentication.

GET /api/v1/admin/config

List all configuration entries grouped by category.

Response 200 OK

[
  {
    "key": "rate_limit_requests_per_minute",
    "value": 60,
    "source": "default",
    "editable": true,
    "category": "Rate Limiting",
    "description": "Maximum requests per user per minute.",
    "value_type": "int",
    "enum_options": null,
    "min_value": 1,
    "max_value": 10000
  }
]

Field	Type	Description
`key`	string	Configuration key name
`value`	string \| int \| bool	Current resolved value
`source`	string	Resolution source: `db`, `env`, or `default`
`editable`	bool	Whether the key accepts API updates
`category`	string	UI grouping category
`description`	string	Human-readable description
`value_type`	string	Expected type: `string`, `int`, `bool`, or `enum`
`enum_options`	string[] \| null	Allowed values for enum-type keys
`min_value`	int \| null	Minimum allowed value (int keys only)
`max_value`	int \| null	Maximum allowed value (int keys only)

Sensitive values (database URLs, passwords) are masked as ***.

PUT /api/v1/admin/config/{key}

Update a single configuration value. Only editable: true keys accept updates.

Path parameters

Parameter	Type	Description
`key`	string	Configuration key to update (lowercase, e.g. `rate_limit_requests_per_minute`)

Request body

{ "value": 120 }

Field	Type	Required	Description
`value`	string \| int \| bool	yes	New value — must match the key’s `value_type` and pass constraint validation

Response 200 OK — the updated ConfigEntry object.

Error codes

Status	Condition
`400 Bad Request`	Key is read-only, value fails type validation, or value is outside min/max bounds
`404 Not Found`	Unknown configuration key

Every config update emits an audit log entry recording the old value, new value, and the acting admin’s identity.

Never-exposed keys

The following keys exist in the settings but are never returned by the config API. They are set exclusively through environment variables or secrets backends:

jwt_secret_key, database_url_sync, audit_webhook_url, audit_webhook_token, cors_origins, security_csp_policy

Configuration keys by category

Application

Read-only keys reflecting the deployment identity.

Key	Env Var	Type	Default	Editable	Description
`app_name`	`APP_NAME`	string	`Arbitex`	no	Application display name
`app_version`	`APP_VERSION`	string	`0.1.0`	no	Application semantic version
`debug`	`DEBUG`	bool	`false`	no	Whether debug mode is enabled
`database_url`	`DATABASE_URL`	string	(masked)	no	Database connection URL (masked for security)
`jwt_algorithm`	`JWT_ALGORITHM`	string	`HS256`	no	JWT signing algorithm

Security

Authentication and access control settings.

Key	Env Var	Type	Default	Editable	Constraints	Description
`jwt_access_token_expire_minutes`	`JWT_ACCESS_TOKEN_EXPIRE_MINUTES`	int	`60`	yes	5–1440	JWT access token expiration in minutes
`security_hsts_enabled`	`SECURITY_HSTS_ENABLED`	bool	`true`	yes	—	Send Strict-Transport-Security header
`prompt_guard_enabled`	`PROMPT_GUARD_ENABLED`	bool	`true`	yes	—	Toggle prompt injection defenses
`prompt_guard_log_level`	`PROMPT_GUARD_LOG_LEVEL`	enum	`warning`	yes	`debug`, `info`, `warning`, `error`	Log level for prompt guard warnings
`admin_recovery_enabled`	`ADMIN_RECOVERY_ENABLED`	bool	`true`	yes	—	Allow admins to trigger account recovery for users
`self_service_recovery_enabled`	`SELF_SERVICE_RECOVERY_ENABLED`	bool	`false`	yes	—	Allow users to request self-service account recovery via email
`allow_open_registration`	`ALLOW_OPEN_REGISTRATION`	bool	`false`	yes	—	Allow open user registration without an invite token
`max_concurrent_sessions`	`MAX_CONCURRENT_SESSIONS`	int	`5`	yes	1–100	Maximum concurrent sessions per user

Rate Limiting

Key	Env Var	Type	Default	Editable	Constraints	Description
`rate_limit_requests_per_minute`	`RATE_LIMIT_REQUESTS_PER_MINUTE`	int	`60`	yes	1–10000	Maximum requests per user per minute
`rate_limit_admin_exempt`	`RATE_LIMIT_ADMIN_EXEMPT`	bool	`false`	yes	—	Whether admin users bypass rate limits
`usage_throttle_rpm_factor`	`USAGE_THROTTLE_RPM_FACTOR`	string	`0.5`	yes	—	RPM reduction factor when org is in 95–100% usage band (0.0–1.0)

Audit

Key	Env Var	Type	Default	Editable	Constraints	Description
`default_summarizer_model`	`DEFAULT_SUMMARIZER_MODEL`	string	`claude-sonnet-4-20250514`	yes	—	Default model ID used for summarize mode
`audit_sinks`	`AUDIT_SINKS`	string	`db`	yes	Allowed: `db`, `jsonl`, `webhook`	Active audit sinks (comma-separated)
`audit_compression_level`	`AUDIT_COMPRESSION_LEVEL`	int	`1`	yes	1–19	Zstd compression level for audit prompt/response text (1=fastest, 19=max ratio)
`audit_shard_threshold`	`AUDIT_SHARD_THRESHOLD`	int	`10000000`	yes	100000–100000000	Row count threshold per audit shard before creating the next shard
`audit_retention_days`	`AUDIT_RETENTION_DAYS`	int	`90`	yes	7–3650	Days to retain audit shards before dropping

Conversation

Key	Env Var	Type	Default	Editable	Description
`auto_title_enabled`	`AUTO_TITLE_ENABLED`	bool	`true`	yes	Enable automatic title generation for new conversations
`auto_title_model`	`AUTO_TITLE_MODEL`	string	`claude-haiku-3-20240307`	yes	Model for auto-title generation (format: `provider/model_id` or `model_id`)

DLP

Key	Env Var	Type	Default	Editable	Constraints	Description
`dlp_entity_thresholds`	`DLP_ENTITY_THRESHOLDS`	string	`{}`	yes	Valid JSON	Per-entity-type confidence thresholds (JSON). Overrides defaults in the DLP pipeline
`dlp_ner_force`	`DLP_NER_FORCE`	enum	`auto`	yes	`auto`, `always`, `never`	NER detection override: `auto` (pack-based gating), `always` (force NER on), `never` (force NER off). Per-org override via `dlp_ner_force:{org_id}`
`dlp_context_window_chars`	`DLP_CONTEXT_WINDOW_CHARS`	int	`50`	yes	10–200	Character window (±N) around regex matches for context keyword scanning
`image_ocr_enabled`	`IMAGE_OCR_ENABLED`	bool	`false`	yes	—	Enable image text extraction via OCR for DLP scanning. Per-org override via `org.<org_id>.image_ocr_enabled`
`image_ocr_provider`	`IMAGE_OCR_PROVIDER`	enum	`tesseract`	yes	`tesseract`, `azure`, `disabled`	OCR provider. Per-org override via `org.<org_id>.image_ocr_provider`
`image_ocr_max_file_size_mb`	`IMAGE_OCR_MAX_FILE_SIZE_MB`	int	`20`	yes	1–100	Maximum file size in MB for OCR processing
`image_ocr_timeout_seconds`	`IMAGE_OCR_TIMEOUT_SECONDS`	int	`30`	yes	5–120	Per-file OCR timeout in seconds
`image_visual_detection_enabled`	`IMAGE_VISUAL_DETECTION_ENABLED`	bool	`false`	yes	—	Enable visual pattern detection (credit cards, ID cards, QR codes). Per-org override via `org.<org_id>.image_visual_detection_enabled`
`image_visual_sensitivity`	`IMAGE_VISUAL_SENSITIVITY`	enum	`medium`	yes	`low`, `medium`, `high`	Sensitivity for visual pattern detection. Per-org override via `org.<org_id>.image_visual_sensitivity`
`tier0_prompt_injection_enabled`	`TIER0_PROMPT_INJECTION_ENABLED`	bool	`true`	yes	—	Enable Tier 0 TF-IDF prompt injection pre-filter
`tier0_confidence_threshold`	`TIER0_CONFIDENCE_THRESHOLD`	string	`0.95`	yes	—	Confidence threshold for Tier 0 pre-filter. Above this threshold, text is blocked without running the full DLP pipeline

Drift Detection

Key	Env Var	Type	Default	Editable	Constraints	Description
`drift_baseline_window_days`	`DRIFT_BASELINE_WINDOW_DAYS`	int	`7`	yes	1–90	Rolling baseline window in days for drift detection
`drift_comparison_window_minutes`	`DRIFT_COMPARISON_WINDOW_MINUTES`	int	`60`	yes	1–1440	Comparison window in minutes for drift detection
`drift_latency_threshold`	`DRIFT_LATENCY_THRESHOLD`	string	`2.0`	yes	—	Z-score threshold for latency drift anomaly flagging
`drift_topic_threshold`	`DRIFT_TOPIC_THRESHOLD`	string	`0.3`	yes	—	Jensen-Shannon divergence threshold for topic drift anomaly flagging
`drift_behavior_sensitivity`	`DRIFT_BEHAVIOR_SENSITIVITY`	enum	`medium`	yes	`low`, `medium`, `high`	Behavior anomaly detection sensitivity (low=3σ, medium=2σ, high=1.5σ)

Bias & Fairness

Key	Env Var	Type	Default	Editable	Constraints	Description
`fairness_audit_enabled`	`FAIRNESS_AUDIT_ENABLED`	bool	`true`	yes	—	Enable fairness violation audit event emission
`fairness_baseline_days`	`FAIRNESS_BASELINE_DAYS`	int	`7`	yes	1–90	Rolling baseline window in days for fairness analysis
`fairness_group_attributes`	`FAIRNESS_GROUP_ATTRIBUTES`	string	`department,role`	yes	—	Comma-separated user metadata attributes to compare for bias detection
`fairness_sentiment_threshold`	`FAIRNESS_SENTIMENT_THRESHOLD`	string	`0.15`	yes	—	Sentiment deviation threshold for flagging bias (0.01–1.0)

Evaluation

Model intelligence and evaluation settings.

Key	Env Var	Type	Default	Editable	Constraints	Description
`eval_comparison_enabled`	`EVAL_COMPARISON_ENABLED`	bool	`true`	yes	—	Enable model comparison evaluation feature
`eval_max_prompts_per_comparison`	`EVAL_MAX_PROMPTS_PER_COMPARISON`	int	`50`	yes	1–200	Maximum prompts per model comparison run
`eval_analytics_trend_days`	`EVAL_ANALYTICS_TREND_DAYS`	int	`30`	yes	1–365	Default number of days for analytics trend data
`eval_health_weights`	`EVAL_HEALTH_WEIGHTS`	string	`{"latency":0.2,"quality":0.2,"safety":0.2,"fairness":0.2,"evaluation":0.2}`	yes	Valid JSON, must sum to 1.0	Weights for health score computation per signal

Agent Execution

Key	Env Var	Type	Default	Editable	Constraints	Description
`agent_enabled`	`AGENT_ENABLED`	bool	`false`	yes	—	Enable the autonomous agent execution framework
`agent_max_concurrent_sessions`	`AGENT_MAX_CONCURRENT_SESSIONS`	int	`3`	yes	1–20	Maximum concurrent agent sessions per tenant
`agent_max_steps`	`AGENT_MAX_STEPS`	int	`20`	yes	1–100	Maximum steps per agent session
`agent_step_timeout_seconds`	`AGENT_STEP_TIMEOUT_SECONDS`	int	`30`	yes	5–300	Timeout in seconds for each agent step

Quotas

Default per-user quota limits applied when no explicit user or group quota is configured.

Key	Env Var	Type	Default	Editable	Constraints	Description
`default_daily_token_limit`	`DEFAULT_DAILY_TOKEN_LIMIT`	int	`1000000`	yes	0–100000000	Default daily token limit per user (0 = unlimited)
`default_monthly_token_limit`	`DEFAULT_MONTHLY_TOKEN_LIMIT`	int	`30000000`	yes	0–1000000000	Default monthly token limit per user (0 = unlimited)
`default_daily_request_limit`	`DEFAULT_DAILY_REQUEST_LIMIT`	int	`1000`	yes	0–1000000	Default daily request limit per user (0 = unlimited)
`default_monthly_request_limit`	`DEFAULT_MONTHLY_REQUEST_LIMIT`	int	`30000`	yes	0–10000000	Default monthly request limit per user (0 = unlimited)
`default_daily_cost_limit_usd`	`DEFAULT_DAILY_COST_LIMIT_USD`	int	`50`	yes	0–100000	Default daily cost limit per user in USD (0 = unlimited)
`default_monthly_cost_limit_usd`	`DEFAULT_MONTHLY_COST_LIMIT_USD`	int	`1500`	yes	0–1000000	Default monthly cost limit per user in USD (0 = unlimited)

Output Quality

Key	Env Var	Type	Default	Editable	Constraints	Description
`output_scoring_enabled`	`OUTPUT_SCORING_ENABLED`	bool	`false`	yes	—	Enable output quality scoring (hallucination, relevance, toxicity, groundedness)
`output_scoring_timeout_ms`	`OUTPUT_SCORING_TIMEOUT_MS`	int	`5000`	yes	1000–30000	Skip output scoring if provider response latency exceeds this threshold (ms)
`quality_alert_threshold_hallucination`	`QUALITY_ALERT_THRESHOLD_HALLUCINATION`	string	`0.3`	yes	—	Alert when avg hallucination score drops below this (0.0–1.0)
`quality_alert_threshold_relevance`	`QUALITY_ALERT_THRESHOLD_RELEVANCE`	string	`0.3`	yes	—	Alert when avg relevance score drops below this (0.0–1.0)
`quality_alert_threshold_toxicity`	`QUALITY_ALERT_THRESHOLD_TOXICITY`	string	`0.3`	yes	—	Alert when avg toxicity score drops below this (0.0–1.0)
`quality_alert_threshold_groundedness`	`QUALITY_ALERT_THRESHOLD_GROUNDEDNESS`	string	`0.3`	yes	—	Alert when avg groundedness score drops below this (0.0–1.0)

AppGuard

Key	Env Var	Type	Default	Editable	Description
`appguard_ml_enabled`	`APPGUARD_ML_ENABLED`	bool	`false`	yes	Enable ML classifier as primary threat scorer (falls back to heuristic if unavailable)
`appguard_ensemble_mode`	`APPGUARD_ENSEMBLE_MODE`	bool	`false`	yes	Run both heuristic and ML classifier, log A/B comparison in response

Email

SMTP delivery settings for transactional emails (account recovery, security notifications, invitations).

Key	Env Var	Type	Default	Editable	Constraints	Description
`smtp_host`	`SMTP_HOST`	string	`""`	yes	—	SMTP server hostname. Empty = dev mode (emails logged, not sent)
`smtp_port`	`SMTP_PORT`	int	`587`	yes	1–65535	SMTP server port (587 for STARTTLS, 465 for SSL)
`smtp_username`	`SMTP_USERNAME`	string	`""`	yes	—	SMTP authentication username
`smtp_password`	`SMTP_PASSWORD`	string	(masked)	yes	—	SMTP authentication password
`smtp_from_address`	`SMTP_FROM_ADDRESS`	string	`[email protected]`	yes	—	Sender address for transactional emails
`smtp_use_tls`	`SMTP_USE_TLS`	bool	`true`	yes	—	Use TLS/STARTTLS for SMTP connection

Per-org overrides

Several DLP and image scanning keys support per-org overrides using the naming convention org.<org_id>.<key_name>. These are stored as additional entries in the system_configs table and take precedence over the global key for the specified org.

Supported per-org keys:

org.<org_id>.image_ocr_enabled
org.<org_id>.image_ocr_provider
org.<org_id>.image_visual_detection_enabled
org.<org_id>.image_visual_sensitivity
dlp_ner_force:<org_id>

Example — enable OCR for a specific org:

curl -X PUT https://gateway.arbitex.ai/api/v1/admin/config/org.550e8400-e29b-41d4-a716-446655440000.image_ocr_enabled \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"value": true}'